1. Your face is captured in a photo or on video
2. FRS reads the geometry of your face - 68 different facial features
3. FRS matches your face to a stored image or determines something about you

FRS learns to do this though machine learning and being fed data sets of images

• Security - police, airports
• Authentication - unlocking mobile phones
• Social media - Facebook photo tagging, animoji
• Marketing - smart billboards, targeted advertising
• State control - tracking the Uighur population in China, social credit scores

Microsoft Azure demo

What happens when FRS gets it wrong?

• Joy Buolamwini used a FRS in her project which did not recognize her - only worked when she wore a white mask
• She then ran a study to evaluate how well FRS does at identifying gender in people with different skin colours
• Error rate was 0.8% for white men - 20-35% for black women
• This is due to the data sets fed into the FRS - largely white and male faces

• Loss of anonymity - government can track you anywhere e.g Chinese jaywalkers
• Police have tracked and arrested protesters using FRS
• No legislation to make police in the USA remove your image from their database
• Image databases are vulnerable to hackers
• What can be done with your image?

• GDPR - covers biometric data as a “special category” and specifically mentions “facial images”
• This means you have to have explicit consent or a really good reason to use it
• Biometric Information Privacy Act (Illinois) - less restrictive then GDPR
• Users must be transparent in how they are using FRS - but doesn’t apply to the police
• FRS banned entirely in San Francisco and Oakland

Ban Facial Recognition Map

• Self reporting - Face Recognition Vendor Test

• Algorithmic Justice League - Safe Face Pledge
• Best practices - transparency, testing for fairness and accuracy, asking for consent
• Or maybe take a more personal approach…

Facial Recognition Software should be regulated - but how?

The datasets used to train FRS need to be more inclusive and representative of the population

Legislation ensuring self-reporting of bias and errors for all developers and sellers of FRS

BIPA seems like a good compromise to address privacy issues while allowing wider use